Quick Answer
Among object-detection models from labs with a published paper or leaderboard record, Microsoft’s Table Transformer for Table Structure Recognition is the top pick as of September 2026 because it shares the newest listed release date and wins the downloads tiebreak; ranking note: newest release first, then downloads, with no public benchmark yet for any candidate and the newest available releases included because too few qualify within the last 12 months [1][3][5][6]. In order, the ranking is Table Transformer for Table Structure Recognition (Microsoft), Table Transformer for Table Detection (Microsoft), DEtection TRansformer (DETR) with ResNet-50 (Facebook), and DEtection TRansformer (DETR) with ResNet-101 (Facebook) [1][3].
Key Takeaways
- Ranking note: eligibility is limited to object detection models from labs with a published paper or leaderboard record, using Hugging Face’s
object-detectionpipeline tag.[2][4] Too few recent releases qualify, so the newest available candidates are listed despite their Hugging Face publication dates falling in 2022.[1][3][5][6] No public benchmark scores any candidate; the ordering is newest release first, then downloads, with downloads used only to break ties.[1][3][5][6] - Table Transformer for Table Structure Recognition (Microsoft) ranks first because its Hugging Face publication date ties Table Transformer for Table Detection (Microsoft) at 2022-10-14, while its monthly downloads are higher: 1,057,033 versus 530,864.[3][5] Its 29M parameters imply 14.5 MB of weight memory at 4-bit, estimated (params x 0.5 bytes); its license is MIT.[3]
- Table Transformer for Table Detection (Microsoft) ranks second under that date-and-download ordering.[3][5] Its 29M parameters imply 14.5 MB of weight memory at 4-bit, estimated (params x 0.5 bytes), and its license is MIT.[5]
- DEtection TRansformer (DETR) with ResNet-50 (Facebook) ranks third: its 2022-03-02 Hugging Face publication date ties DEtection TRansformer (DETR) with ResNet-101 (Facebook), but monthly downloads are 349,004 versus 17,800.[1][6] Its 42M parameters imply 21 MB of weight memory at 4-bit, estimated (params x 0.5 bytes); its license is Apache-2.0.[1]
- DEtection TRansformer (DETR) with ResNet-101 (Facebook) ranks fourth under the same ordering.[1][6] Its 61M parameters imply 30.5 MB of weight memory at 4-bit, estimated (params x 0.5 bytes); its license is Apache-2.0.[6]
- CPU performance remains unranked: every candidate has “no public benchmark yet” for this comparison, and each lists a context length of 1K tokens.[1][3][5][6] The weight-memory estimates above do not establish total runtime RAM requirements or measured CPU speed.
How do these object detection models compare on parameters, estimated weight memory, context, licenses and public benchmarks?
| Model | Org | Params | Quant/VRAM | Released (date) | License | Key benchmark (date) |
|---|---|---|---|---|---|---|
| Table Transformer for Table Structure Recognition [3] | Microsoft [3] | 29M [3] | 4-bit weights: 14.5 MB estimated (params x 0.5 bytes), using 29M params; runtime memory not published [3] | 2022-10-14 [3] | MIT [3] | no public benchmark yet [3] |
| Table Transformer for Table Detection [5] | Microsoft [5] | 29M [5] | 4-bit weights: 14.5 MB estimated (params x 0.5 bytes), using 29M params; runtime memory not published [5] | 2022-10-14 [5] | MIT [5] | no public benchmark yet [5] |
| DEtection TRansformer (DETR) with ResNet-50 [1] | Facebook [1] | 42M [1] | 4-bit weights: 21 MB estimated (params x 0.5 bytes), using 42M params; runtime memory not published [1] | 2022-03-02 [1] | Apache-2.0 [1] | no public benchmark yet [1] |
| DEtection TRansformer (DETR) with ResNet-101 [6] | Facebook [6] | 61M [6] | 4-bit weights: 30.5 MB estimated (params x 0.5 bytes), using 61M params; runtime memory not published [6] | 2022-03-02 [6] | Apache-2.0 [6] | no public benchmark yet [6] |
Which object detection models should you consider for a CPU-only PC?
1. Table Transformer for Table Structure Recognition
Table Transformer for Table Structure Recognition by Microsoft ranks first because it shares the newest listed Hugging Face publication date and wins the download tiebreak.[3][5][1][6] The ordering is newest release first, then downloads.[3][5][1][6] The ranking admits only object detection models from labs with a published paper or leaderboard record, using the Hugging Face object-detection pipeline tag.[2][4] The family has too few recent releases, so the newest available candidates are listed; Microsoft’s entry was first published on Hugging Face on October 14, 2022.[3]
The model has 29 million parameters, a listed context length of 1K tokens, and F32 weights occupying 0.1 GB.[3] Its MIT license permits use under that license’s terms.[3] At 4-bit precision, weight storage would be approximately 14.5 MB, estimated (params x 0.5 bytes) from the cited 29 million parameters.[3] That calculation describes hypothetical weight storage, rather than a verified quantized download or total runtime memory requirement.
For local deployment on a CPU-only PC, allow RAM for the inference runtime and image processing alongside the weights; a specific RAM minimum or CPU speed recommendation cannot be established here. The intended use is table structure recognition in document extraction workflows.[3][4] The caveat is performance uncertainty: no public benchmark yet. Its ranking therefore establishes neither detection accuracy nor CPU inference speed, and the weight estimate does not demonstrate practical CPU support for quantized inference.
2. Table Transformer for Table Detection
Table Transformer for Table Detection by Microsoft ranks second because its October 14, 2022 Hugging Face publication date ties with the structure-recognition variant, which wins the download tiebreak.[5][3] Ranking note: newest release first, then downloads; too few recent releases are available, so the newest available candidates are included.[1][3][5][6] Eligibility is limited to object detection models from labs with a published paper or leaderboard record, with the Hugging Face pipeline tag object-detection.[2][4]
Microsoft lists 29M parameters, a context length of 1K tokens, F32 weights occupying 0.1 GB, and an MIT license.[5] The model has no public benchmark yet, so its position does not establish detection accuracy or CPU speed. The practical use case is table detection within a document-processing workflow, consistent with Microsoft’s “PubTables-1M: Towards comprehensive table extraction from unstructured documents” paper.[4] Choose it when locating tables is the task.
For local hardware planning, weights alone would occupy 14.5 MB at 4-bit—estimated (params x 0.5 bytes), using the listed 29M parameters.[5] That calculation does not establish total runtime RAM requirements or confirm a working quantized CPU implementation. A specific minimum RAM configuration cannot be recommended from weight size alone. The caveat for a CPU-only PC is unverified runtime performance: check CPU support and measure memory use and latency in your intended runtime before committing to deployment.
3. DEtection TRansformer (DETR) with ResNet-50
DEtection TRansformer (DETR) with ResNet-50 by Facebook ranks third under the ordering rule: newest release first, then downloads.[1][3][5][6] The ranking admits only object detection models from labs with a published paper or leaderboard record, using Hugging Face’s object-detection pipeline tag. The shortlist uses the newest available entries because too few qualify within the requested release window. Its Hugging Face publication date is March 2, 2022; download counts break the tie with Facebook’s DEtection TRansformer (DETR) with ResNet-101.[1][6]
The model has 42 million parameters, lists a context length of 1K tokens, and provides F32 weights totaling 0.2 GB under the Apache-2.0 license.[1] Weight storage at 4-bit precision is approximately 21 MB, estimated (params x 0.5 bytes) from the cited 42 million parameters.[1] That calculation describes weight storage alone; it does not establish quantized runtime support or the total RAM needed for local inference. A CPU-only PC’s practical memory requirement must be checked with the intended runtime and workload.
Consider this model for a local object-detection evaluation when the Apache-2.0 license suits your project and you can measure performance on your own PC.[1] The benchmark status is “no public benchmark yet” for this ranking. The practical caveat is that its position does not demonstrate CPU speed or detection accuracy: publication dates and download counts determine the order, so validate latency and output quality before committing.
4. DEtection TRansformer (DETR) with ResNet-101
DEtection TRansformer (DETR) with ResNet-101 by Facebook ranks fourth because its publication date ties the other Facebook entry, while its download count is lower.[1][6] The ordering is newest release first, then downloads; no public benchmark scores any candidate. Eligibility is limited to object detection models from labs with a published paper or leaderboard record and the Hugging Face pipeline tag object-detection. The family has too few recent releases, so the newest available candidates are included.[1][3][5][6]
Facebook lists 61M parameters, a context length of 1K tokens, F32 weights occupying 0.2 GB, and an Apache-2.0 license.[6] The model was first published on Hugging Face on March 2, 2022, and recorded 17,800 downloads over the preceding 30 days as of September 25, 2026.[6] Its architectural foundation is Facebook’s “End-to-End Object Detection with Transformers.”[2] Benchmark status: no public benchmark yet.
For local use, consider the model for object detection experiments where you can measure performance on your own CPU. At 61M parameters,[6] 4-bit weight storage would be 30.5 MB, estimated (params x 0.5 bytes).[6] Treat that calculation as a storage estimate, not a validated quantized deployment or a total RAM requirement. The practical caveat is hardware uncertainty: no CPU latency measurement or validated RAM requirement is established, so the small weight footprint alone cannot demonstrate responsive inference.
Why does the ranking use newest release first, then downloads, when no public benchmark scores any candidate?
The ranking uses newest release first, then downloads, because no public benchmark scores any candidate, leaving no shared score that can justify a performance-based order.[1][3][5][6] Downloads serve only as a tiebreak between models with the same publication date; popularity does not establish detection accuracy or CPU speed.
The ranking admits only object detection models from labs with a published paper or leaderboard record, using Hugging Face’s object-detection pipeline tag.[1][2][3][4][5][6] The intended release window has too few candidates, so the newest available models are listed despite their older publication dates.[1][3][5][6]
Microsoft’s Table Transformer for Table Structure Recognition (microsoft/table-transformer-structure-recognition) takes the first position because its publication date ties the newest listed release and its download count breaks that tie.[3][5] Microsoft’s Table Transformer for Table Detection (microsoft/table-transformer-detection) shares the October 14, 2022 publication date, with 530,864 downloads against the structure-recognition model’s 1,057,033 downloads over the reported thirty-day period.[3][5]
Facebook’s Detection Transformer with ResNet-50 (facebook/detr-resnet-50) precedes Facebook’s Detection Transformer with ResNet-101 (facebook/detr-resnet-101): both were published on March 2, 2022, while their reported thirty-day downloads are 349,004 and 17,800, respectively.[1][6]
Every entry therefore carries “no public benchmark yet.”[1][3][5][6] Read the order as a release-and-popularity ordering, not a measured CPU performance result. Parameter counts, weight sizes and licenses remain useful comparison details, but none substitutes for a shared benchmark or determines placement.
How should you interpret estimated weight memory when planning CPU-only deployment?
Treat estimated weight memory as a planning baseline for parameter storage, not a complete RAM requirement or a prediction of CPU performance.
Microsoft’s Table Transformer Structure Recognition (microsoft/table-transformer-structure-recognition) has 29M parameters [3], giving 14.5 MB of estimated weight memory at 4-bit precision (estimated (params x 0.5 bytes)) [3]. Microsoft’s Table Transformer Detection (microsoft/table-transformer-detection) has the same 29M parameter count [5] and the same 14.5 MB estimate (estimated (params x 0.5 bytes)) [5].
Facebook’s Detection Transformer (DETR) with ResNet-50 (facebook/detr-resnet-50) has 42M parameters [1], giving 21 MB at 4-bit precision (estimated (params x 0.5 bytes)) [1]. Facebook’s Detection Transformer (DETR) with ResNet-101 (facebook/detr-resnet-101) has 61M parameters [6], giving 30.5 MB at 4-bit precision (estimated (params x 0.5 bytes)) [6].
Use those calculations as hypothetical weight budgets. Do not treat them as confirmation that a compatible quantized checkpoint exists or that your chosen CPU runtime can execute it. Check the actual checkpoint format and runtime support before committing to a deployment.
Budget separately for image processing, intermediate tensors, runtime allocations and the surrounding application. Validate peak RAM and inference latency with representative images and your intended concurrency. A parameter-based estimate alone cannot establish whether a model fits your available RAM or meets your response-time target.
Why does the shortlist include older releases from labs with a published paper or leaderboard record?
The shortlist includes older releases because the eligible families have too few recent releases, so the newest available checkpoints are included as an explicit exception to the recency gate.[1][3][5][6] Their inclusion should not be read as evidence of a recent model launch or measured CPU performance.
The ranking admits only object detection models from labs with a published paper or leaderboard record, using Hugging Face’s object-detection pipeline tag. Microsoft’s Table Transformer models accompany PubTables-1M: Towards comprehensive table extraction from unstructured documents.[4] Facebook’s DETR models accompany End-to-End Object Detection with Transformers.[2] Those papers establish a publication record; they do not establish a CPU performance ranking.
Ranking note: no public benchmark scores any candidate, so the ordering is newest release first, then downloads.[1][3][5][6] Microsoft’s Table Transformer for structure recognition takes first place because it shares the latest listed publication date with Microsoft’s Table Transformer for table detection and has the higher download count.[3][5] Facebook’s DETR with ResNet-50 follows, then Facebook’s DETR with ResNet-101.[1][6] Each entry carries “no public benchmark yet”; download counts serve only as a tiebreak.
For CPU-only planning, Microsoft’s structure-recognition checkpoint has 29M parameters, giving approximately 14.5 MB of four-bit weight storage, estimated (params x 0.5 bytes).[3] That calculation covers weights alone. The shortlist does not establish quantized runtime support, total RAM requirements, or CPU latency, so its ordering should guide inspection rather than imply measured speed.
Frequently Asked Questions
Which object detection model ranks first for a CPU-only PC?
Microsoft’s Table Transformer Structure Recognition (microsoft/table-transformer-structure-recognition) ranks first because it shares the newer publication date and wins the download tiebreak.[3][5] Microsoft’s Table Transformer Detection (microsoft/table-transformer-detection) shares that publication date.[5] The family has too few recent releases, so the newest available candidates are listed; their Hugging Face publication dates fall outside the requested recent-release window.[1][3][5][6] The position does not establish superior CPU performance.
How are the object detection models ranked?
Ranking note: no public benchmark scores any candidate, so the ordering is newest release first, then downloads.[1][3][5][6] Scope: this ranking admits only object detection models from labs with a published paper or leaderboard record (Hugging Face pipeline tags: object-detection).[2][4] The order is Table Transformer Structure Recognition,[3] Table Transformer Detection,[5] Facebook’s DEtection TRansformer (DETR) ResNet-50 (facebook/detr-resnet-50),[1] then Facebook’s DEtection TRansformer (DETR) ResNet-101 (facebook/detr-resnet-101).[6]
How much memory would quantized weights need?
For Table Transformer Structure Recognition, 29M parameters imply 14.5 MB, estimated (params x 0.5 bytes).[3] Table Transformer Detection also has 29M parameters: 14.5 MB, estimated (params x 0.5 bytes).[5] DETR ResNet-50 has 42M parameters: 21 MB, estimated (params x 0.5 bytes).[1] DETR ResNet-101 has 61M parameters: 30.5 MB, estimated (params x 0.5 bytes).[6] Each estimate covers weights alone; deployment memory and quantization support remain unestablished.
Which model has demonstrated better CPU performance?
No CPU performance winner is established for this candidate set. Each entry carries the status “no public benchmark yet” for the ranking.[1][3][5][6] Publication dates and download counts determine placement, but neither measures inference latency. A practical comparison would need the intended processor, runtime and workload held consistent. Treat CPU speed as an unresolved deployment question when choosing among these models.
What weight sizes and context lengths are listed?
Table Transformer Structure Recognition lists F32 weights of 0.1 GB and a context length of 1K tokens.[3] Table Transformer Detection lists the same specifications.[5] DETR ResNet-50 lists F32 weights of 0.2 GB and 1K tokens,[1] as does DETR ResNet-101.[6] Those weight figures describe the published weights; they do not establish total application RAM requirements or measured CPU performance.
What licenses do these object detection models use?
Microsoft’s Table Transformer Structure Recognition uses the MIT license,[3] and Table Transformer Detection also uses MIT.[5] Facebook’s DETR ResNet-50 uses Apache-2.0,[1] as does DETR ResNet-101.[6] For deployment planning, record the license alongside the exact model identifier and review its terms before distribution. License choice is a separate consideration from the release-date ordering and the unresolved CPU performance comparison.
Sources
- facebook/detr-resnet-50 model card (Hugging Face) — 2026-09-25
- End-to-End Object Detection with Transformers — 2020-05-26
- microsoft/table-transformer-structure-recognition model card (Hugging Face) — 2026-09-25
- PubTables-1M: Towards comprehensive table extraction from unstructured documents — 2021-09-30
- microsoft/table-transformer-detection model card (Hugging Face) — 2026-09-25
- facebook/detr-resnet-101 model card (Hugging Face) — 2026-09-25